HKVM-RAG hypergraph method boosts multi-hop RAG accuracy by up to 11 F1

Multi-hop RAG tasks—answering questions that require connecting multiple pieces of evidence—have long been limited by how retrieved text is organized. Traditional dense retrievers score passages independently, while graph-based methods often rely on pairwise or entity-centered keys that fragment the evidence chain. In a new paper submitted to ICDE 2027, researchers Mingyu Zhang and Ying Ma introduce HKVM-RAG (Key-Value-Separated Hypergraph Evidence Organization for Multi-Hop RAG). The method assembles answer-path hyperedges from cached passage-level LLM evidence tuples, using them as retrieval keys while retaining passage text as answer values. This key-value separation allows the system to maintain strong retrieval signals without losing the original context.

On standard benchmarks, HKVM-RAG delivers significant gains: +3.426 F1 on 2WikiMultiHopQA and +3.592 F1 on MuSiQue compared to the KG-PPR baseline. When combined with a dense-aware controller using frozen ColBERTv2, it achieves 88.846, 65.073, and 85.810 F1 on 2WikiMultiHopQA, MuSiQue, and HotpotQA respectively—improvements of +11.084, +6.763, and +5.966 F1 over ColBERTv2 alone. The authors note that HotpotQA reveals higher structured support coverage doesn't always yield standalone answer gains, suggesting HKVM works best as an evidence-control signal rather than a dense-retrieval replacement. Source-level ablations confirm that non-hypergraph structured signals fail to match these gains.